Topaz DeNoise AI

Overview

Conventional noise reduction has one fundamental problem, which is that it cannot tell noise from detail. It smooths everything above a threshold, so the grain goes and so does the eyelash, the fabric weave and the texture of the wall behind the subject. The result is a clean image that looks like plastic, and every photographer has thrown away frames rather than deliver one.

This works differently. The model was trained on matched pairs of noisy and clean images, so it has learned what real detail looks like underneath noise and reconstructs it rather than averaging it away. On a frame shot at very high sensitivity, feathers, hair, stitching and text survive a pass that removes the grain almost entirely, which is the difference between a usable image and a deleted one.

Several models are available because different noise has different character. The standard model covers most cases, a low light model handles the combination of noise and underexposure that comes from pushing shadows several stops, a severe noise model handles the extreme end, and a raw model works on the file before demosaicing where noise reduction is more effective than it is after. Comparison view shows the models side by side on the same crop so the choice is made by eye.

It runs as a standalone application and as a plugin inside the common hosts, so it can sit in an existing workflow as a step rather than a destination. Batch processing pushes a setting across a folder for a full shoot, output preserves the original bit depth and colour space, and the sharpening and detail recovery controls sit alongside the noise reduction so the two are balanced against each other rather than applied in separate passes that fight.

Key Features

• Trained noise reduction that reconstructs detail rather than averaging it away
• Multiple models covering standard, low light, severe noise and raw file processing
• Comparison view showing several models side by side on the same crop
• Raw processing before demosaicing where noise reduction is most effective
• Detail recovery and sharpening balanced against the noise reduction in one pass
• Automatic setting estimation with full manual override
• Batch processing across a folder with a saved setting
• Standalone application and plugin operation inside the common hosts
• Original bit depth and colour space preserved on output
• Graphics accelerated processing with a software fallback

What Changed in This Build

• Improved detail retention on fine texture such as hair and fabric
• Faster processing on current generation graphics hardware
• Better handling of colour noise in deep shadow areas
• Raw model extended to cover recent camera bodies

System Requirements

Processor	Intel or AMD 64-bit with AVX2 support
Memory	8 GB minimum, 16 GB recommended
Graphics	GPU with 4 GB VRAM recommended, 2 GB minimum
Storage	3 GB free
Display	1920 x 1080 or higher

Release Details

Full title	Topaz DeNoise AI
Version string	3.7.2
Publisher	Topaz Labs
Category	Photo Editing
Licence	Full version, no term limit
Platform	Windows 10 and 11, 64-bit
Interface language	English
Archive size	1.18 GB
Mirrors	4 active, all reporting online
Listed	refreshed 3 months ago, checked again 40 minutes ago

Installation Notes

• Apply noise reduction early in the edit, before sharpening and heavy local adjustment, rather than at the end.
• Judge the result at 100 percent. Noise reduction that looks perfect at fit to screen frequently is not.
• Use the raw model where the workflow allows it, since working before demosaicing gives a visibly better result.

Before You Start

Read the notes above before running setup. Most of the problems people write in about are covered there, and almost all of them come down to an older version of the same software still being installed, a graphics driver that predates the release, or a security suite quarantining part of the package mid install.

There is no archive password on this or any other entry in the library. If something you downloaded elsewhere claims to come from here and asks for one, it did not come from here. Nothing on this site is behind a survey, a shortener or a countdown either.

If the entry does not match what the page describes, say so on the contact page and include the version string from the release details table. Reports naming a specific version get checked first because the checker can reproduce them straight away.

Frequently Asked